2026-06-23

Claude Code’s “Extended Thinking” Isn’t the Actual Reasoning

For Claude Code users and people building agent workflows on top of Claude, this is a surprisingly important distinction: what looks like a reasoning trace may not be the raw chain of thought you assumed it was. If you’re treating those logs as an audit trail, a debugging artifact, or evidence of why an agent acted, the gap between “summary” and “actual thinking” really matters.

Key Points

The author inspected Claude Code session logs and found that the “thinking blocks” were not readable reasoning text, but a long signature.
The article says Claude encrypts its reasoning into that signature, and Anthropic holds the key.
The local machine does not receive the key, so the machine cannot decrypt the full thinking output.
The API returns a summary of reasoning, not the reasoning itself.
Getting the full thinking output requires an enterprise agreement, according to the source.
The author argues that this matters if you want a trustworthy record of an agent’s logic during a session.
The post compares the behavior to saving a JPEG as a BMP, then presenting it as if it were the original JPEG: the conversion loses information.
The docs are described as indirect, and the warning is that it’s easy to miss that “extended thinking” is a summary rather than the full process.
The author also notes that logs can capture inputs, outputs, and actions, but not the actual internal reasoning that drove the model’s behavior.

My Take

What strikes me is that this is less a revelation about Claude specifically and more a reminder that “reasoning visibility” is a slippery concept in LLM products. I think a lot of developers intuitively read “thinking” output as a literal transcript of the model’s internal deliberation, when in practice it may be a mediated artifact with missing information.

I’m sympathetic to the complaint here. If you’re using Claude Code in a serious workflow, especially for compliance, debugging, or incident review, it would be easy to over-trust those logs. I’d be careful about promising anyone an audit trail based on local files alone, because that sounds stronger than what the source says the system actually provides.

At the same time, I think there’s a useful product story hidden in this mess: even a summary can still be helpful for understanding intent, spotting bad reasoning patterns, or making agent behavior less opaque. But I’d want Anthropic to be much more explicit about the limits. The complaint about the docs being indirect feels fair; if a feature is summarized reasoning, say that plainly.

If I were using Claude Code, I’d log the inputs, outputs, tool calls, and state changes myself, and treat any “extended thinking” display as supplemental context rather than ground truth. I’d also be curious whether Anthropic will make the boundary between summary and actual reasoning clearer over time, because I think that clarity matters more than the cleverness of the feature itself.

The takeaway: don’t mistake Claude Code’s extended thinking display for a complete record of the model’s internal reasoning. It may still be useful, but it is not the same thing as an authentic, fully recoverable thinking trace.